Nvidia has unveiled Alpamayo 2 Edge, a vision-language-action model designed to plan vehicle paths and provide real-time explanations for driving decisions. The model was announced on October 21, 2026, at GTC Berlin. It features 2 billion parameters and is a lightweight version of the 34-billion-parameter Alpamayo 2 Super, optimized through distillation.
The model operates on the Nvidia DRIVE AGX Thor platform, refreshing driving plans at a rate of 10 Hz. Following optimization with TensorRT, it achieves a pixel-to-trajectory latency of 160 to 190 milliseconds. Research indicates that the use of closed-loop reinforcement learning during training improved the model's performance by 48.9%.
Nvidia is releasing the model on Hugging Face under permissive licensing, with additional resources for customization available on GitHub. The Alpamayo 2 Edge is part of a suite targeting Level 4 autonomous vehicles and robotaxis. Simultaneously, the company introduced the Halos SAFE safety framework, developed in collaboration with dSPACE, IVEX, and TÜV Rheinland.
Volvo Cars announced its adoption of the Nvidia Hyperion compute system and the Alpamayo stack for future vehicle development. Additionally, Nvidia plans to launch the Alpamayo Agentic Factory in early 2027, which will serve as a blueprint for integrating AI agents into automotive systems.
